Google AI Overview Miscounts Letters in 4 Basic Word Tests as Token Design Limits Spelling
Updated
Updated · TechCrunch · May 28
Google AI Overview Miscounts Letters in 4 Basic Word Tests as Token Design Limits Spelling
5 articles · Updated · TechCrunch · May 28
Google’s AI Overview failed basic letter-counting prompts, claiming two Ps in “Google,” one r in “poop,” and misspelling “journalism” and Trump while answering.
Google said counting letters inside words is a known weakness and blamed the issue on large language models’ token-based design, which encodes chunks of text rather than reading letters like humans.
Researchers cited by TechCrunch said that architecture makes spelling unusually hard to fix because models operate on tokens that may represent words, syllables or letters, not stable word-by-word structure.
The errors add to earlier AI Overview stumbles, including last week’s patched “disregard” definition glitch and past answers that cited satire and suggested eating rocks or putting glue on pizza.
The episode underscores a broader limit of generative AI: systems that can solve complex tasks can still fail at elementary text handling, making output verification essential.
Why does Google's advanced AI fail at spelling tasks even a child can master?
Can AI's foundational flaws be fixed, or will we always need to second-guess its answers?
As AI summaries reduce web traffic, what is the future for online content creators and publishers?